Two-stage image decomposition and color regulator for low-light image enhancement

Low-lighting is a common condition in data collection due to environmental restrictions. However, high-level pattern recognition tasks such as object detection require the datasets to be more clear. Thus, low-light image enhancement is necessary. Noise and color distortion are two major problems of the existing enhancement algorithms. This paper has proposed a low-light image enhancement algorithm that integrates denoising and color restoration. First, we propose a two-stage hybrid decomposition network, which can perform modified Retinex-decomposition on paired images, and then extract principal components of the decomposed low-light images to handle the nonlinear residuals, thereby obtaining reliable reflectance and illumination maps. Then, in order not to over-smooth the details and edges of the image, we use a flexible joint function to train the hybrid network. Finally, we create a color regulator in the HSI (Hue-Saturation-Intensity) space to correct the distortion in RGB space caused by coupling between pixels. Experimental results on public datasets show that the proposed method greatly enhanced the quality of low-light images.